27th local time.OpenAI announced the training of a GPT-4 based program called CriticGPT The model for finding the ChatGPT ChatbotsErrors in the outputIt can write comments to emphasize inaccuracies in the answers generated by ChatGPT. It is possible to write comments highlighting inaccuracies in the answers generated by ChatGPT.
CriticGPT is described as being designed to assist human AI trainers with their work -- using a technology called "Reinforcing Learning from Human Feedback(Note: Reinforcement Learning from Human Feedback, RLHF)" technique to train and improve GPT-4 responses.
However, as ChatGPT becomes more accurate, the errors become more insidious, making the AI trainer's job more and more "difficult." OpenAI explains that this is one of the fundamental limitations of RLHF -- the model gradually becomes more and more difficult to use than anyone else who can provide feedback. Anyone who can provide feedbackmore knowledgeable, model harmonization may become increasingly difficult with it.
Currently, when CriticGPT attempts to answer from ChatGPT'sSpot the error.OpenAI points out that real-world errors can be spread all over the answer to a question.many parts, which is something CriticGPT will need to address in the future. "Our focus is on being able to point out errors in one place, but in the future we will need to address decentralized errors as well."